Microsoft
Microsoft · AI / ML · Filed Jul 20, 2026 · USPTO published Nov 12, 2026 · This week's release
Microsoft Files Patent for Letting AI Fill In a Video's Missing Detail
Microsoft's patent has an AI guess what a sharp video looked like from a blurry copy, then sends a small patch file to fix its mistakes.
- Publication number
- US 2026/0361899 A1(US20260361899A1)
- Applicant
- Microsoft Technology Licensing, LLC
- Filed · Published
- Jul 20, 2026 · Nov 12, 2026
- Inventors
- Ziyad A. IBRAHIM, Sudhakar V. PRABHU
- Status when we published
- Waiting for an examiner
- Document
- 20 claims
- US classification
- 375/240.12
- Parent application
- Continuation of 18/421,809 (filed Jan 24, 2024)(18421809)
What Microsoft's AI video rebuilding does
Ever tried to send a long vacation video and watched the upload bar crawl for an hour? Video files are huge, and shrinking them usually costs you sharpness.
Microsoft's patent describes a different trade. First, the video is shrunk to a lower-quality copy, and that small copy is what gets stored or sent. On the other end, an AI model looks at the small copy, predicts what the sharp original most likely looked like, and rebuilds it.
The AI will not guess perfectly, so the sender checks its work. It compares the AI's rebuilt version to the original and packs the leftover differences into a small extra file. Your device combines the two and gets back something close to the original.
… converting the first version to a second version of the input media, the second version having a second resolution; compressing the second version according to a media compression format, thereby producing compressed data for the second version; …
Translation: The system shrinks a high-resolution video down to a lower resolution before standard compression.
How the AI predicts a video's compressed code
This filing covers a video compression pipeline built around a generative AI model (the same family of software that powers chatbots, here trained on video). The sender shrinks the original video to a lower-resolution version, with fewer pixels, a lower frame rate or lower quality, and compresses it with an ordinary video format. The main example is H.264.
The AI does not read raw pixels. It reads partially decompressed data: the compressed file is opened just far enough to pull out its syntax elements (the built-in instructions of a video file, such as how blocks of image moved between frames). Each one becomes a token (a labeled piece, like a word for a language model), and the conversion can be reversed exactly.
From there:
- The AI's encoder (its reading half) takes in the tokens for the low-quality version.
- Its decoder (its writing half) predicts, one token at a time, the tokens for the high-quality version.
- A standard video decoder turns those predictions into frames.
- The sender compares those frames with the original and compresses the differences into an "augmentation stream" (an add-on correction file).
Training used pairs of low and high versions made with matching settings. The model was scored on whether its output follows the video format's rules, and on objective quality measures such as PSNR and SSIM (standard scores for how close two images look).
The decompression tool provides the input tokens to the generative AI model and receives predicted tokens from the generative AI model. The predicted tokens represent output syntax elements, respectively, of compressed data for output media.
Translation: An AI model takes the compressed data and predicts the missing details to rebuild the video.
Why smaller video files matter for streaming
If this works as described, your video could travel as a small low-quality file plus a small correction file, instead of one big high-quality file. The filing notes the low-quality copy can have a much lower bitrate (the amount of data per second of video), and that is what drives storage and streaming costs.
The filing says the compress and decompress steps could run on a laptop or phone, while training happens on cloud servers. It also says the same idea could apply to audio and images, and to video formats beyond H.264. No measured savings are given, so the size of the win is an open question.
Microsoft's ninth filing we've tracked in AI photo editing since July builds on earlier applications like sharpening blurry video and rewriting spoken words in recordings.
Source: US 2026/0361899 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Microsoft patent news breaks.
Add Patentlyze on Google
One click.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings · 14 sheets
14 drawing sheets from US 2026/0361899 A1 · click any drawing to enlarge
Show the other 6 sheets
More from Microsoft
All 646 Microsoft patents →
More in AI image & video
All 223 AI image & video patents →
Adobe
Adobe Files Patent for AI That Adds Missing Edges to Your Photos With One Click
Google
Google Files Patent for Fixing Blurry Phone Photos With Less Battery Drain
Adobe
Adobe Patents AI Fill That Compares Colors Twice to Reduce Patchy Results
Nvidia
No comments yet